Skip to content

GitLab

Commit 421b718d authored by rocking's avatar rocking
Browse files

Merge commit '57fadf6f' into gemm_layernorm_welford

parents 2cf6f30b 57fadf6f
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1668 additions and 136 deletions
example/22_cgemm/cgemm_xdl_int8.cpp
View file @ 421b718d
......@@ -117,7 +117,7 @@ int main(int argc, char* argv[])
exit(0);
}
return run_cgemm_xdl<ADataType,
return !run_cgemm_xdl<ADataType,
BDataType,
CDataType,
ALayout,
......
example/24_batched_gemm/CMakeLists.txt 0 → 100644
View file @ 421b718d
add_custom_target(example_batched_gemm_xdl)
add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
add_dependencies(example_batched_gemm_xdl
example_batched_gemm_xdl_fp32
example_batched_gemm_xdl_fp16
example_batched_gemm_xdl_bfp16
example_batched_gemm_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)
endif()
example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp 0 → 100644
View file @ 421b718d
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = BF16;
using BDataType = BF16;
using AccDataType = F32;
using CShuffleDataType = BF16;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
#include "run_batched_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
example/24_batched_gemm/batched_gemm_xdl_fp16.cpp 0 → 100644
View file @ 421b718d
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F16;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on
#include "run_batched_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
example/24_batched_gemm/batched_gemm_xdl_fp32.cpp 0 → 100644
View file @ 421b718d
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F32;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>;
// clang-format on
#include "run_batched_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
example/24_batched_gemm/batched_gemm_xdl_int4.cpp 0 → 100644
View file @ 421b718d
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = ck::int4_t;
using BDataType = ck::int4_t;
using AccDataType = int32_t;
using CShuffleDataType = int32_t;
using DsDataType = ck::Tuple<>;
using EDataType = ck::int4_t;
using KernelADataType = int8_t;
using KernelBDataType = int8_t;
using KernelEDataType = int8_t;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
// clang-format off
< ALayout, //ALayout
BLayout, //BLayout
DsLayout, //DsLayout
ELayout, //ELayout
KernelADataType, //ADataType
KernelBDataType, //BDataType
AccDataType, //AccDataType
CShuffleDataType, //CShuffleDataType
DsDataType, //DsDataType
KernelEDataType, //EDataType
AElementOp, //AElementwiseOperation
BElementOp, //BElementwiseOperation
CDEElementOp, //CDEElementwiseOperation
GemmDefault, //GEMMSpecialization
1, // NumGemmKPrefetchStage
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXdl
32, // NPerXdl
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransfer ThreadCluster Lengths_K0_M_K1
S<1, 0, 2>, // ABlockTransfer ThreadCluster ArrangeOrder
S<1, 0, 2>, // ABlockTransfer SrcAccessOrder
2, // ABlockTransfer SrcVectorDim
16, // ABlockTransfer SrcScalarPerVector
16, // ABlockTransfer DstScalarPerVector_K1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransfer ThreadCluster Lengths_K0_N_K1
S<1, 0, 2>, // BBlockTransfer ThreadCluster ArrangeOrder
S<1, 0, 2>, // BBlockTransfer SrcAccessOrder
2, // BBlockTransfer SrcVectorDim
16, // BBlockTransfer SrcScalarPerVector
16, // BBlockTransfer DstScalarPerVector_K1
1, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 64, 1, 4>, // CBlockTransferClusterLengths_MBlock_MWaveMPerXdl_NBlock_NWaveNPerXdl
16>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
#define BUILD_INT4_EXAMPLE
#include "run_batched_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
example/24_batched_gemm/batched_gemm_xdl_int8.cpp 0 → 100644
View file @ 421b718d
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = int8_t;
using BDataType = int8_t;
using AccDataType = int32_t;
using CShuffleDataType = int8_t;
using DsDataType = ck::Tuple<>;
using EDataType = int8_t;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>;
// clang-format on
#include "run_batched_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
example/24_batched_gemm/run_batched_gemm_example.inc 0 → 100644
View file @ 421b718d
#include <random>
#pragma once
struct ProblemSize final
{
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t stride_A = K;
ck::index_t stride_B = K;
ck::index_t stride_C = N;
ck::index_t batch_stride_A = M * K;
ck::index_t batch_stride_B = K * N;
ck::index_t batch_stride_C = M * N;
ck::index_t batch_count = 16;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
using namespace ck::literals;
#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
static_assert(sizeof(ADataType) == sizeof(KernelADataType));
static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
static_assert(sizeof(EDataType) == sizeof(KernelEDataType));
#endif
auto& [M,
N,
K,
stride_A,
stride_B,
stride_C,
batch_stride_A,
batch_stride_B,
batch_stride_C,
batch_count] = problem_size;
// GEMM shape
auto f_host_tensor_descriptor = [](std::size_t batch_count_,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
Tensor<BDataType> b_g_k_n(
f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
#ifdef BUILD_INT4_EXAMPLE
Tensor<KernelEDataType> e_g_m_n_device_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
#else
Tensor<EDataType> e_g_m_n_device_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
#endif
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
switch(config.init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpaceSize());
#ifdef BUILD_INT4_EXAMPLE
const Tensor<KernelADataType> a_g_m_k_converted(a_g_m_k);
const Tensor<KernelBDataType> b_g_k_n_converted(b_g_k_n);
a_device_buf.ToDevice(a_g_m_k_converted.mData.data());
b_device_buf.ToDevice(b_g_k_n_converted.mData.data());
#else
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
#endif
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
// do GEMM
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
{},
c_device_buf.GetDeviceBuffer(),
M,
N,
K,
batch_count,
stride_A,
stride_B,
{},
stride_C,
batch_stride_A,
batch_stride_B,
{},
batch_stride_C,
a_element_op,
b_element_op,
cde_element_op);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
if(config.do_verification)
{
c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
using ReferenceBatchedGemmInstance =
ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
BDataType,
EDataType,
AccDataType,
AElementOp,
BElementOp,
CDEElementOp>;
auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
Tensor<EDataType> e_g_m_n_host_result(
f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
auto ref_argument = ref_batched_gemm.MakeArgument(
a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
ref_invoker.Run(ref_argument);
#ifdef BUILD_INT4_EXAMPLE
const Tensor<EDataType> e_device_result_converted(e_g_m_n_device_result);
pass &= ck::utils::check_err(e_device_result_converted.mData, e_g_m_n_host_result.mData);
#else
pass = ck::utils::check_err(
e_g_m_n_device_result.mData, e_g_m_n_host_result.mData, "Error: Incorrect results c");
#endif
}
if(config.time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t flop = std::size_t(2) * batch_count * M * N * K;
std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
sizeof(BDataType) * batch_count * K * N +
sizeof(EDataType) * batch_count * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
}
return pass ? 0 : 1;
}
bool run_batched_gemm_example(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
std::mt19937 gen(11939);
std::uniform_int_distribution<int> dis(0, 15);
problem_size.M = 256 * (dis(gen) + 1);
problem_size.N = 128 * (dis(gen) + 1);
problem_size.K = 64 * (dis(gen) + 2);
problem_size.stride_A = problem_size.K;
problem_size.stride_B = problem_size.K;
problem_size.stride_C = problem_size.N;
problem_size.batch_stride_A = problem_size.M * problem_size.K;
problem_size.batch_stride_B = problem_size.K * problem_size.N;
problem_size.batch_stride_C = problem_size.M * problem_size.N;
problem_size.batch_count = 16;
if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
exit(0);
}
return run_batched_gemm(problem_size, config);
}
example/24_batched_gemm_e_permute/CMakeLists.txt deleted 100644 → 0
View file @ 2cf6f30b
add_example_executable(example_batched_gemm_e_permute_xdl_fp16 batched_gemm_e_permute_xdl_fp16.cpp)
example/30_grouped_convnd_fwd_bias_relu_add/CMakeLists.txt
View file @ 421b718d
add_example_executable(example_grouped_convnd_fwd_bias_relu_add_xdl_fp16 grouped_convnd_fwd_bias_relu_add_xdl_fp16.cpp)
target_link_libraries(example_grouped_convnd_fwd_bias_relu_add_xdl_fp16 PRIVATE utility)
add_example_executable(example_grouped_convnd_fwd_bias_relu_add_xdl_fp32 grouped_convnd_fwd_bias_relu_add_xdl_fp32.cpp)
target_link_libraries(example_grouped_convnd_fwd_bias_relu_add_xdl_fp32 PRIVATE utility)
add_example_executable(example_grouped_convnd_fwd_bias_relu_add_xdl_bf16 grouped_convnd_fwd_bias_relu_add_xdl_bf16.cpp)
target_link_libraries(example_grouped_convnd_fwd_bias_relu_add_xdl_bf16 PRIVATE utility)
add_example_executable(example_grouped_convnd_fwd_bias_relu_add_xdl_int8 grouped_convnd_fwd_bias_relu_add_xdl_int8.cpp)
target_link_libraries(example_grouped_convnd_fwd_bias_relu_add_xdl_int8 PRIVATE utility)
\ No newline at end of file
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_convnd_fwd_bias_relu_add_xdl_int4 grouped_convnd_fwd_bias_relu_add_xdl_int4.cpp)
endif() # USE_BITINT_EXTENSION_INT4
example/30_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_common.hpp
View file @ 421b718d
......@@ -26,13 +26,16 @@ void print_helper_msg()
}
template <ck::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename InKernelDataType,
typename WeiKernelDataType,
typename CShuffleDataType,
typename OutDataType,
typename OutKernelDataType,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp,
typename InUserDataType,
typename WeiUserDataType,
typename OutUserDataType,
typename DeviceConvNDFwdInstance>
int run_grouped_conv_fwd_bias_relu_add(bool do_verification,
int init_method,
......@@ -47,12 +50,12 @@ int run_grouped_conv_fwd_bias_relu_add(bool do_verification,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{
Tensor<InDataType> in(in_g_n_c_wis_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
Tensor<OutDataType> bias(bias_g_n_k_wos_desc);
Tensor<OutDataType> residual(residual_g_n_k_wos_desc);
Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
Tensor<InUserDataType> in(in_g_n_c_wis_desc);
Tensor<WeiUserDataType> wei(wei_g_k_c_xs_desc);
Tensor<OutUserDataType> bias(bias_g_n_k_wos_desc);
Tensor<OutUserDataType> residual(residual_g_n_k_wos_desc);
Tensor<OutUserDataType> out_host(out_g_n_k_wos_desc);
Tensor<OutKernelDataType> out_device(out_g_n_k_wos_desc);
std::cout << "in: " << in.mDesc << std::endl;
std::cout << "wei: " << wei.mDesc << std::endl;
......@@ -64,26 +67,38 @@ int run_grouped_conv_fwd_bias_relu_add(bool do_verification,
{
case 0: break;
case 1:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
bias.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
in.GenerateTensorValue(GeneratorTensor_2<InUserDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiUserDataType>{-5, 5});
bias.GenerateTensorValue(GeneratorTensor_2<OutUserDataType>{-5, 5});
break;
default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
bias.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
in.GenerateTensorValue(GeneratorTensor_3<InUserDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiUserDataType>{-0.5, 0.5});
bias.GenerateTensorValue(GeneratorTensor_3<OutUserDataType>{-0.5, 0.5});
}
DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
DeviceMem bias_device_buf(sizeof(OutDataType) * bias.mDesc.GetElementSpaceSize());
DeviceMem residual_device_buf(sizeof(OutDataType) * residual.mDesc.GetElementSpaceSize());
DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
DeviceMem in_device_buf(sizeof(InKernelDataType) * in.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiKernelDataType) * wei.mDesc.GetElementSpaceSize());
DeviceMem bias_device_buf(sizeof(OutKernelDataType) * bias.mDesc.GetElementSpaceSize());
DeviceMem residual_device_buf(sizeof(OutKernelDataType) * residual.mDesc.GetElementSpaceSize());
DeviceMem out_device_buf(sizeof(OutKernelDataType) * out_device.mDesc.GetElementSpaceSize());
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
const Tensor<InKernelDataType> in_converted(in);
const Tensor<WeiKernelDataType> wei_converted(wei);
const Tensor<OutKernelDataType> bias_converted(bias);
const Tensor<OutKernelDataType> residual_converted(residual);
in_device_buf.ToDevice(in_converted.mData.data());
wei_device_buf.ToDevice(wei_converted.mData.data());
bias_device_buf.ToDevice(bias_converted.mData.data());
residual_device_buf.ToDevice(residual_converted.mData.data());
#else // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
in_device_buf.ToDevice(in.mData.data());
wei_device_buf.ToDevice(wei.mData.data());
bias_device_buf.ToDevice(bias.mData.data());
residual_device_buf.ToDevice(residual.mData.data());
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
......@@ -154,7 +169,7 @@ int run_grouped_conv_fwd_bias_relu_add(bool do_verification,
float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = conv_param.GetFlops();
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
std::size_t num_btype = conv_param.GetByte<InUserDataType, WeiUserDataType, OutUserDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / avg_time;
......@@ -168,8 +183,8 @@ int run_grouped_conv_fwd_bias_relu_add(bool do_verification,
Tensor<CShuffleDataType> c_host(out_g_n_k_wos_desc);
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
InDataType,
WeiDataType,
InUserDataType,
WeiUserDataType,
CShuffleDataType,
InElementOp,
WeiElementOp,
......@@ -196,10 +211,22 @@ int run_grouped_conv_fwd_bias_relu_add(bool do_verification,
out_device_buf.FromDevice(out_device.mData.data());
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
const Tensor<OutUserDataType> out_device_converted(out_device);
return ck::utils::check_err(out_device_converted.mData,
out_host.mData,
"Error: incorrect results!",
1e-5f,
1e-4f)
? 0
: 1;
#else // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
return ck::utils::check_err(
out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f)
? 0
: 1;
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
}
return 0;
......
example/30_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_bf16.cpp
View file @ 421b718d
......@@ -7,13 +7,19 @@
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
using InDataType = ck::bhalf_t;
using WeiDataType = ck::bhalf_t;
// kernel data types
using InKernelDataType = ck::bhalf_t;
using WeiKernelDataType = ck::bhalf_t;
using AccDataType = float;
using CShuffleDataType = float;
using BiasDataType = ck::bhalf_t;
using ResidualDataType = ck::bhalf_t;
using OutDataType = ck::bhalf_t;
using BiasKernelDataType = ck::bhalf_t;
using ResidualKernelDataType = ck::bhalf_t;
using OutKernelDataType = ck::bhalf_t;
// tensor data types
using InUserDataType = InKernelDataType;
using WeiUserDataType = WeiKernelDataType;
using OutUserDataType = OutKernelDataType;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
......@@ -40,12 +46,12 @@ using DeviceGroupedConvNDFwdInstance =
WeiLayout,
ck::Tuple<BiasLayout, ResidualLayout>,
OutLayout,
InDataType,
WeiDataType,
InKernelDataType,
WeiKernelDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<BiasDataType, ResidualDataType>,
OutDataType,
ck::Tuple<BiasKernelDataType, ResidualKernelDataType>,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
......@@ -181,13 +187,16 @@ int main(int argc, char* argv[])
});
return run_grouped_conv_fwd_bias_relu_add<1,
InDataType,
WeiDataType,
InKernelDataType,
WeiKernelDataType,
CShuffleDataType,
OutDataType,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
InUserDataType,
WeiUserDataType,
OutUserDataType,
DeviceGroupedConvNDFwdInstance<1,
InLayout,
WeiLayout,
......@@ -290,13 +299,16 @@ int main(int argc, char* argv[])
});
return run_grouped_conv_fwd_bias_relu_add<2,
InDataType,
WeiDataType,
InKernelDataType,
WeiKernelDataType,
CShuffleDataType,
OutDataType,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
InUserDataType,
WeiUserDataType,
OutUserDataType,
DeviceGroupedConvNDFwdInstance<2,
InLayout,
WeiLayout,
......@@ -413,13 +425,16 @@ int main(int argc, char* argv[])
});
return run_grouped_conv_fwd_bias_relu_add<3,
InDataType,
WeiDataType,
InKernelDataType,
WeiKernelDataType,
CShuffleDataType,
OutDataType,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
InUserDataType,
WeiUserDataType,
OutUserDataType,
DeviceGroupedConvNDFwdInstance<3,
InLayout,
WeiLayout,
......
example/30_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_fp16.cpp
View file @ 421b718d
......@@ -7,13 +7,19 @@
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
// kernel data types
using InKernelDataType = ck::half_t;
using WeiKernelDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using BiasDataType = ck::half_t;
using ResidualDataType = ck::half_t;
using OutDataType = ck::half_t;
using BiasKernelDataType = ck::half_t;
using ResidualKernelDataType = ck::half_t;
using OutKernelDataType = ck::half_t;
// tensor data types
using InUserDataType = InKernelDataType;
using WeiUserDataType = WeiKernelDataType;
using OutUserDataType = OutKernelDataType;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
......@@ -40,12 +46,12 @@ using DeviceGroupedConvNDFwdInstance =
WeiLayout,
ck::Tuple<BiasLayout, ResidualLayout>,
OutLayout,
InDataType,
WeiDataType,
InKernelDataType,
WeiKernelDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<BiasDataType, ResidualDataType>,
OutDataType,
ck::Tuple<BiasKernelDataType, ResidualKernelDataType>,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
......@@ -181,13 +187,16 @@ int main(int argc, char* argv[])
});
return run_grouped_conv_fwd_bias_relu_add<1,
InDataType,
WeiDataType,
InKernelDataType,
WeiKernelDataType,
CShuffleDataType,
OutDataType,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
InUserDataType,
WeiUserDataType,
OutUserDataType,
DeviceGroupedConvNDFwdInstance<1,
InLayout,
WeiLayout,
......@@ -290,13 +299,16 @@ int main(int argc, char* argv[])
});
return run_grouped_conv_fwd_bias_relu_add<2,
InDataType,
WeiDataType,
InKernelDataType,
WeiKernelDataType,
CShuffleDataType,
OutDataType,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
InUserDataType,
WeiUserDataType,
OutUserDataType,
DeviceGroupedConvNDFwdInstance<2,
InLayout,
WeiLayout,
......@@ -413,13 +425,16 @@ int main(int argc, char* argv[])
});
return run_grouped_conv_fwd_bias_relu_add<3,
InDataType,
WeiDataType,
InKernelDataType,
WeiKernelDataType,
CShuffleDataType,
OutDataType,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
InUserDataType,
WeiUserDataType,
OutUserDataType,
DeviceGroupedConvNDFwdInstance<3,
InLayout,
WeiLayout,
......
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View file @ 421b718d
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment